How I Used Stable Diffusion to Create a Fully Generated Photoshoot for My T-Shirt Merch

Hey there! I’m Georgiy Ermakov and I am a designer.
In my work I love experimenting with different design techniques and approaches.

Recently, I’ve been tinkering with Stable Diffusion, and like many others, I decided to train the model on images of myself. But instead of creating endless versions of myself in medieval capes and Superman suits, I wanted to take a more practical approach.

I thought, what if I could use my digital self to model real products?

Before we dive in, let me show you what I look like in real life:

Hello there!

First, I started experimenting with different methods of training.
Textual Inversion and Lora didn’t yield the results I was hoping for. Most of the generated images were deformed and strange-looking versions of the original training set.

I believe the above mentioned methods can produce good results, but it just didn’t work for me in this particular case

I also decided to skip the Hypernetwork approach this time, as I heard it produces similar results to Textual Inversion.

I moved on to Dreambooth training since I’ve heard it gives better results. After some trial and error and several models and trainings, I started to see some promising results. Although they weren’t perfect, some of the generated images looked quite interesting and consistent.

So after some more tweaking and finetuning with different models, I finally hit the sweet spot. Now we are talking!

Although I understand that it is a more prettified and polished version of the real-life me, I totally don’t mind since this is exactly what is needed for a fashion photoshoot. I call him Alter Geo

Having the consistent result from the trained model I first tried it on some imaginary futuristic clothing concepts.

Not bad it looks!

Or some photoshoot for a new Depeche Mode album… Just kidding

So finally, after I had the pipeline figured out, it was time to try it for the real product: My merch t-shirt with a local design meme

It is a little joke, kind of opposite to the “New Visual Culture” slogan. Roughly can be translated as “New Visual Unculture” or “New Visual Barbarism”

For this “photoshoot” I decided to try various styles, including more realistic, gritty, film grain candid photography. After some generating and polishing, I ended up with several images.

The T-Shirt text is added later in Photoshop. Currently, I couldn’t achieve normal text rendering through training.

Although the “person” in the image looks a little inconsistent from image to image in general it looks pretty cool. Also, it allows for saving quite a lot of money on photographers and traveling

Hello, Elon!

In conclusion, I can say that I found this approach very flexible and useful. You can create almost any image or photo concept easily.

Pros & Cons

• Convenient. You can train on any model or any object and have it in your digital library after that
• Fast. You can get thousands of possible images in an hour or even less
• Flexible. You can try different styles and approaches easily
• Not easy. It was quite hard to train the model to get consistently good results from it
• Needs tweaking. The generated images are low-res, so if you need them to be printed you will need to do some upscaling. This can be done with various instruments, including Stable Diffusion addons, but it is not a one-button solution for now if you want production-quality results

Please note! The images above were processed in Photoshop for color correction and polishing of the details, but all the main elements were fully generated in Stable Diffusion, except for the text on the t-shirt

BTW, you can order the T-Shirt from Amazon here: US, UK, DE, JP

Or If you want to support my creative work, you can buy me a tea!

My other projects and links:

• Neurosofia.com — AI Generated Postcards about Sofia, Bulgaria
• https://twitter.com/eerrmmkkvv
• https://www.instagram.com/ermkv.world/